Tim O’Reilly’s scenario planning post on AI and jobs1 is worth reading. It’s careful, honest about uncertainty, and the four-quadrant framework is a useful thinking tool. But there’s a structural assumption embedded in the design that’s worth pulling out.
O’Reilly builds his framework around two dimensions: scale and speed of adoption, and whether AI is used for efficiency or to ‘do more’—solve new problems, serve unmet demand, create new categories. The upper quadrants, augmentation and transformation, are available by strategic choice. Companies that want to inhabit them can. The robust strategy, he argues, is to aim for the upper right.
But look at where LLMs work well: coding assistance, document drafting, summarisation, search augmentation. Or with agentic AI: customer support triage, code review pipelines, document processing workflows. There’s value here, but notice the shape of that list—it hasn’t kept expanding in surprising directions. The map of where LLMs work has been getting clearer, not more open-ended. That’s the pattern of a nail gun, not electrification.
Nail guns transformed construction productivity, but nobody built buildings that were impossible before them. The technology settled into the cost curve and stayed there. Electrification reorganised what was possible—it didn’t just make existing things cheaper, it created new categories, new industries, new relationships between capital and labour.
If LLMs are nail guns, O’Reilly’s ‘do more’ axis doesn’t collapse because firms make bad choices. The technology’s nature forecloses it. The upper quadrants aren’t a strategic choice companies make or decline to make—they’re simply not on the table.
We’ve had two sequential hype cycles in eighteen months. Late 2025 the shine came off LLMs. Early 2026 was the pivot to agentic AI. Now the same pattern is emerging again, as the data on the limitations of agentic AI become piles up. The failure modes are no longer surprising—they’re systematic and repeatable. Sequential pivots are themselves data. Not bad luck, nor growing pains, but a signal.
This changes how we read the evidence O’Reilly marshals. The PwC wage premium for AI-skilled workers, the Vanguard finding that AI-exposed occupations are growing—these are consistent with nail gun dynamics. A nail gun creates real wage premiums for the workers who use them well. It concentrates productivity gains. It reshapes who does what within an industry. None of that requires it to be electrification.
O’Reilly’s relational sector thesis and the Imas demand-elasticity argument are both working hard to rescue the upper quadrants. They’re interesting, but assume AI resembles electricity—that it reorganises what’s possible rather than just what’s cheap. If LLMs are nail guns, the relational sector still grows, but through a weaker mechanism: the income effect of rising efficiency getting spent somewhere, rather than AI actively unlocking new human capacity. That’s historically slower. It also distributes the transition pain differently: more concentrated, longer duration, less legible as transformation while it’s happening.
That leaves a simpler, less exciting framework than O’Reilly’s. Not four quadrants but one spectrum: how fast does efficiency-seeking adoption proceed, and who captures the gains? The interesting strategic questions aren’t about which quadrant to aim for. They’re concerned with positioning on the cost curve, and the political economy of who the productivity gains flow to.
O’Reilly’s closing argument—that the future is up to us, that every firm choosing efficiency over expansion compounds toward a hollowed economy—is right as a moral claim and useful as competitive advice. A firm that uses AI to do new things will outcompete one that just cuts headcount, even in a nail-gun world. But it doesn’t change the macro trajectory if the technology itself doesn’t open new possibility space.
The hype cycle pattern is evidence, not proof. Two sequential pivots in eighteen months could still be growing pains rather than a signal. The honest position is that we don’t yet have the data to be certain—our measurement tools aren’t fine-grained enough to see what’s actually happening in the labour market, as O’Reilly himself acknowledges. But the burden of proof has shifted. The electricity analogy needs to show its work.
I’ve written about the structural limits of LLMs, the specific failure modes of agentic AI, and how we might determine whether LLMs are more like nail guns or electrification—the short version being that the limitations aren’t engineering problems awaiting better models, they’re structural features of how language models work. That’s the underlying argument for why the nail gun reading is more than just scepticism about hype: it’s grounded in what the technology actually is.
- O’Reilly, Tim. “Scenario Planning for AI and the ‘Jobless Future.’” O’Reilly Radar, April 20, 2026. https://www.oreilly.com/radar/scenario-planning-for-ai-and-the-jobless-future/. ↩︎